Yesterday I proved NB2 can render text at any complexity level. Today I wanted to know if it understands context.

I gave NB2 the same character and asked it to make book covers across five genres.

It didn't just change the background. It changed the art style. The typography. The costume. The facial expression. The color palette. Even the author name's gender.

Twenty-four images later, I have proof that NB2 has something I'm calling genre intelligence. The ability to understand book cover conventions as a complete system, not just a visual style.

And then I broke it on purpose to prove the point.

One character. One reference image. Five genres: Thriller, Fantasy, Romance, Sci-Fi, and Self-Help. Each got a genre-appropriate prompt with a unique title. Then Variation F: a thriller title paired with romance visuals. The "what NOT to do" control.

This is the headline finding. NB2 didn't apply filters or color shifts to the same base rendering. It used fundamentally different artistic approaches per genre.

Thriller. Photorealistic noir. Rain-slicked cobblestones. Neon reflections. Desaturated blue palette. The character looks back over his shoulder into fog. NB2 gave the title a shattered glass overlay effect, a visual metaphor for "psychological" that I never prompted.

Fantasy. Painterly digital illustration. Mountains, castle on cliff, storm clouds breaking into golden sunset. The character got leather armor, a crystal-tipped prismatic staff, and rune stones at his feet. The title rendered in ornate metallic with a holographic rainbow fill. This doesn't look like the same model that made the thriller.

Romance. Soft, dreamy photography. Golden hour on a beach. Pastel pink. Floating rose petals. The character's expression shifted from mischievous to genuinely tender. A soft smile, a hand placed on his heart. The title rendered in elegant brown script. This is the same creature that was lurking in a dark alley two generations ago.

Sci-Fi. Photorealistic CGI. Space station corridor with holographic data panels. Cyan and orange lighting. Deep vanishing point perspective. Clean futuristic sans-serif typography. The holographic displays read "PRISM ARRAY." NB2 built in-world technology from the character's name.

Self-Help. Clean corporate graphic design. Navy-to-white gradient. Bold sans-serif title. One image put the character in a tailored navy suit with a prismatic pocket square that matched his iridescent skin. Another rendered as a physical hardcover book mockup with visible spine shadow. That mockup behavior is worth remembering. It comes back later this week.

NB2 doesn't just dress the character differently. It decides how much to change based on genre distance from reality.

Romance is contemporary. Set in the real world. Zero costume changes. The black t-shirt and red pants stayed.

Thriller is also real-world but darker. One image got a leather jacket. The rest kept the reference outfit.

Sci-Fi is futuristic but still grounded. One image got a military flight jacket. And NB2 turned the character's existing Adobe Firefly branding into military rank patches. It didn't just change the clothes. It genre-adapted the brand itself.

Fantasy is fully fictional. Three out of four images got complete costume redesigns: leather armor with a crystal staff, full adventurer gear with a sword and belt pouch, torn battle-worn clothing. NB2 understood that a t-shirt doesn't belong in a fantasy world.

Self-Help got a navy suit. Because that's what self-help authors wear on their covers.

I never specified author names. I just wrote "author name at bottom." NB2 generated sixteen different fictional author names across the session.

For thriller, it chose male-coded names: Elias Vance. A.J. Reed. J.P. Mercer.

For romance, it chose female-coded names: Anya Reed. Elara Stone. Eliza Vance.

For fantasy, it used gender-neutral initials: A.R. Sterling. C.L. Eldora.

For self-help, it added professional credentials: DR. ALICE LEE.

It even applied publishing conventions. Adding "BY" before the name, "A Novel" as a subtitle, and in one case a lowercase "by" that mirrors real paperback formatting.

The limitation: NB2 has a name generation bias. "Elara Vance" appeared three times across different genres. And about 21% of images showed placeholder text like "(Author Name)" instead of generating one. If you need a specific name, write it in the prompt.

To prove genre conventions matter, I ran Variation F: the thriller title "THE PRISM EFFECT" with romance visuals. Pastel pink, flower petals, golden sunlight, delicate script font.

The genre-correct thriller covers averaged 8.54. The mismatched version averaged 7.44. That's a 1.10 point drop. A clean, measurable penalty for getting the genre wrong.

But the most fascinating result was NB2's attempt to resolve the contradiction. When it couldn't make the visuals match "psychological thriller," it wrote a tagline: "Sometimes the most beautiful surfaces hide the darkest truths."

I didn't prompt that. NB2 invented copy to bridge the gap between what the image showed and what the genre label demanded. That's not text rendering. That's creative problem-solving.

Two other images added "A Psychological Thriller" as a genre label. NB2 couldn't show the genre, so it told it instead. When the system breaks, NB2 tries to fix it with words.

Fantasy led the session at 8.94 average, driven by a 9.45 peak. The armored crystal staff image (B-1) is portfolio quality without editing. Self-Help produced the most useful output format: physical book mockups that authors can share directly on social media. Romance proved that NB2 can transform a mischievous creature into a tender romantic lead through expression alone.

The full ranking:

#1: Fantasy (8.94 avg). Most creative freedom, strongest visual payoff. B-1 hit 9.45, the new session high (beating Monday's 9.25).

#2: Romance (8.61 avg). Highest consistency (9/10 average). The expression transformation is the standout finding.

#3: Thriller (8.54 avg). Tied with Sci-Fi on average, but the shattered glass text treatment gives it the edge in uniqueness.

#4: Sci-Fi (8.54 avg). The in-world naming and brand-as-military-insignia behaviors are remarkable, even if scores don't reflect it.

#5: Self-Help (8.40 avg). Lower scores but highest practical utility. The book mockup format is immediately usable.

#6: Mismatch Control (7.44 avg). The proof. Genre conventions matter. 1.10 points below the genre-correct version.

The data says: NB2 is strongest when the genre gives it the most creative freedom (fantasy) and weakest when genre signals conflict (mismatch control).

The pupil issue and shirt logo dropout from Monday both persisted. New issues specific to book covers:

Placeholder author text appeared in 21% of images. NB2 sometimes outputs "(Author Name)" instead of generating one. Fix: specify the author name in your prompt, or edit in Firefly Boards.

Barefoot character in most images because the reference sheet doesn't include shoes. Add footwear to the reference image or specify in the prompt.

Name repetition. "Elara Vance" is NB2's default author name. If you need variety, specify names.

Reference image override. One romance generation produced a random human character instead of Prism. The reference image influence isn't absolute. Genre conventions can occasionally overpower it. That tension between style and reference becomes a central finding tomorrow.

Here's the genre template structure. Replace the bracketed sections with your content:

Thriller:

Fantasy:

Romance:

Sci-Fi:

Self-Help:

If you're self-publishing or creating covers for content, NB2 through Adobe Firefly understands genre as a system. You don't need to specify "use painterly illustration style" for fantasy or "render in photorealistic noir" for thriller. The genre label itself triggers the right conventions.

The key insight: treat the genre description in your prompt as a design brief, not just a category tag. "Book cover for psychological thriller" activates an entire visual vocabulary. Dark palettes, distressed typography, tense body language, desaturated tones. NB2 knows what a thriller looks like because it knows what thrillers are.

Tomorrow: putting Prism on t-shirts, stickers, and merchandise. If NB2 can do genre intelligence for books, can it do product intelligence for merch?

Testing methodology: Nano Banana 2 (@NanoBanana), a partner model inside Adobe Firefly (@AdobeFirefly). All images scored using a weighted 5-dimension rubric. Minimum 4 generations per variation before drawing conclusions.